SeisGAN: Improving Seismic Image Resolution and Reducing Random Noise Using a Generative Adversarial Network

Seismic images are essential for understanding the subsurface geological structure and resource distribution. However, the accuracy and certainty of geological analysis using seismic images are limited by the resolution and signal-to-noise ratio. Simultaneously improving resolution and suppressing random noise with traditional methods can be quite challenging. This research proposes a new approach called SeisGAN which leverages a generative adversarial network to address the challenge at hand. Due to the lack of high-resolution noiseless and low-resolution noisy seismic data, stochastic parameter control is employed to simulate a vast range of diverse, paired seismic data for SeisGAN training. The results on the synthetic dataset demonstrate that the proposed method is effective in enhancing the resolution and suppressing the random noise in the original images. Spectrum analysis shows that the proposed method increases the bandwidth of the original data, primarily at high frequencies. Ablation experiments reveal that, under similar conditions, SeisGAN outperforms traditional convolutional neural networks. Incorporating the VGG loss in the generator loss function improves the model's ability to recover high-frequency details. The application of the technique on two publicly available field seismic datasets indicates SeisGAN's excellent generalizability, despite being trained only on synthetic seismic data. Compared with bicubic interpolation and traditional noise suppression and resolution enhancement methods, SeisGAN is capable of effectively suppressing the random noise and enhancing the dominant frequency of field seismic data, making it easier to identify adjacent thin layers and fault features, even for small-scale faults. The zoomed images are clearer and easier to interpret. Furthermore, an example of automatic machine fault identification demonstrates the significant contribution of the SeisGAN-enhanced image to accurate fault recognition.